A Robust Object Recognition Using Modified YOLOv5 Neural Network

Abstract

Human beings can recognize and detect the object in any image or video with ease, but when it comes to machines dealing with the same task, it’s quite challenging. In the past years, there has been significant progress in the area of computer vision. However, researchers are still working on enhancing the performance of object recognition algorithms in terms of accuracy and speed. This paper presents one of the greatest CNN representatives, You Only Look Once (YOLO), which tackles object detection in the simplest and most efficient way. This study primarily focuses on object recognition with modified YOLOv5. Here, we have implemented the model on a dataset comprising five objects. We have analyzed various preprocessing and feature extraction techniques to enhance the performance of the proposed model. The highest mean average precision (mAP) achieved by the modified YOLOv5 is 90.7%. The recognition performance of the modified YOLOv5 is compared with the YOLOv4 algorithm. It is also compared with various conventional object detection algorithms outside the YOLO family, such as Support vector machine (SVM), deep Convolutional neural network (CNN), and EfficientNet. The results proved to be promising as compared to the other state-of-the-art techniques.